Language-Invariant Multilingual Speaker Verification for the TidyVoice 2026 Challenge

A research team has introduced a novel approach to multilingual speaker verification (SV), a critical task for security and personalization in voice-activated systems. Their system, designed for the TidyVoice 2026 Challenge, tackles the core problem where traditional SV models perform poorly across languages because speaker embeddings become contaminated with language-specific information. The researchers built their model on the multilingual self-supervised w2v-BERT 2.0 backbone, enhancing it with Layer Adapters and Multi-scale Feature Aggregation to better utilize its complex internal representations.

The key innovation is a language-adversarial training strategy. By implementing a Gradient Reversal Layer, the model is actively forced to learn speaker representations that a secondary classifier cannot use to identify the language being spoken. This promotes truly language-invariant embeddings. To combat data scarcity, the team also employed a multilingual zero-shot text-to-speech system to synthesize speech in multiple languages, artificially expanding their training dataset's linguistic diversity. Experimental results confirmed that fine-tuning the large pre-trained model provided a strong baseline, but the adversarial training further boosted cross-lingual robustness, and synthetic data augmentation delivered significant gains when real training data was limited. The source code is publicly available, facilitating further research and application.